Emneord:
masking læring
Publikasjoner
-
Nielsen, Sebastian Grau
(2026).
Automatic Online Changepoint Detection with Recurrent Neural Networks.
Fulltekst i vitenarkiv
Vis sammendrag
Online changepoint detection is the task of detecting distributional changes in time series data observed in real time. This problem imposes strict demands on the computational complexity of algorithms, since the current observation needs to be processed before the next one arrives. Recently, feedforward neural networks have attracted interest for changepoint detection in the offline (non-sequential) setting. In the online setting, however, these approaches may fail because they are non-sequential and may thus be highly computationally expensive. In this work, we propose a class of recurrent neural networks (RNNs) for online changepoint detection that are inherently sequential, and we show that multiple well-known grid- and window-based statistical tests can be represented as simple RNNs. Leveraging these representations, we present theoretical results that give high-probability bounds on the error rate and detection delay of our trained RNN classifiers. These bounds imply that, with sufficient data, a trained RNN will have a performance that is comparable with these statistical methods when their underlying assumptions are fulfilled. In numerical experiments, we show that our RNNs perform on par with state-of-the-art online statistical algorithms under independent Gaussian noise. Under autocorrelated and heavy-tailed noise, our RNNs appear to significantly outperform these methods. Finally, we present an analysis of the runtimes and show that the runtime of RNN changepoint detection is competitive with existing methods. This is joint work with Idris Eckley, Ingrid Glad, Martin Tveten, Michael Kampffmeyer, and Per August Jarval Moen.
-
Nielsen, Sebastian Grau
(2026).
Automatic Online Changepoint Detection with Recurrent Neural Networks.
Fulltekst i vitenarkiv
Vis sammendrag
Online changepoint detection is the task of detecting distributional changes in time series data observed in real time. This problem imposes strict demands on the computational complexity of algorithms, since the current observation needs to be processed before the next one arrives. Recently, feedforward neural networks have attracted interest for changepoint detection in the offline (non-sequential) setting. In the online setting, however, these approaches may fail because they are non-sequential and may thus be highly computationally expensive. In this work, we propose a class of recurrent neural networks (RNNs) for online changepoint detection that are inherently sequential, and we show that multiple well-known grid- and window-based statistical tests can be represented as simple RNNs. Leveraging these representations, we present theoretical results that give high-probability bounds on the error rate and detection delay of our trained RNN classifiers. These bounds imply that, with sufficient data, a trained RNN will have a performance that is comparable with these statistical methods when their underlying assumptions are fulfilled. In numerical experiments, we show that our RNNs perform on par with state-of-the-art online statistical algorithms under independent Gaussian noise. Under autocorrelated and heavy-tailed noise, our RNNs appear to significantly outperform these methods. Finally, we present an analysis of the runtimes and show that the runtime of RNN changepoint detection is competitive with existing methods. This is joint work with Idris Eckley, Ingrid Glad, Martin Tveten, Michael Kampffmeyer, and Per August Jarval Moen.
Se alle arbeider i NVA
Publisert
27. okt. 2025 11:52
- Sist endret
1. okt. 2026 17:59
